YouTube History (Markdown) Triage

When to use

Use this skill when the user asks to analyze their YouTube watch history and they paste Markdown (instead of relying on Chrome/AppleScript extraction).

Inputs

Required:

• history_markdown: user-pasted Markdown containing YouTube URLs (optionally as [Title](URL) links)

Optional:

• priority_count: how many top picks to recommend (default: 10)
• include_shorts: true|false (default: false)
• focus: 1-sentence focus (default: “AI-assisted software development”)

Outputs

1. A ranked list of candidate videos (top priority_count) with short reasons.
2. A “captured vs not captured” check against raw/ for those candidates.
3. A question asking which videos to capture/distill next.

Procedure

1) Normalize, dedupe, and hard-skip obvious off-topic items

Take the user’s pasted Markdown and run the helper script to:

• extract YouTube video URLs + IDs
• dedupe by video ID
• detect whether each ID is already present in raw/ (best-effort)
• hard-skip obvious off-topic topics (defaults: physics/math, politics/war, sports/fitness, cars, entertainment), configurable via .agents/skills/youtube-history-markdown-triage/filters.json

Run (paste the user’s Markdown into the heredoc):

python3 .agents/skills/youtube-history-markdown-triage/scripts/check_youtube_history_md.py --raw-dir raw --out /tmp/youtube_history_triage.md <<'MD'
<PASTE USER MARKDOWN HERE>
MD

If the pasted history is large, prefer writing output to a file via --out ... to avoid truncation. If the user wants to include shorts, add --include-shorts. Otherwise default behavior is to ignore shorts. If you want to see everything (no filtering), add --no-hard-skip. If you want machine-readable output, add --format json (e.g. --out /tmp/youtube_history_triage.json).

2) Triage for relevance (human judgment)

Using titles + your knowledge of this vault’s focus, classify into:

• suggest: clearly relevant to AI-assisted software development
• maybe: adjacent / uncertain
• skip: unrelated (entertainment, general news without a clear work tie-in, etc.)

Heuristics for suggest:

• coding agents, agent workflows, evals, benchmarking pitfalls
• prompting patterns for software work, specs, tests, refactors, reviews
• developer tooling, IDE workflows, CI/testing, code search, architecture
• org/process changes driven by AI (roles, reviews, quality bars) with concrete detail

Avoid over-weighting hypey titles; prefer topics that plausibly add durable, reusable ideas to the vault.

3) Propose top priorities

Present a ranked top list with:

• title (as given)
• URL
• why it’s high leverage for this vault (1–2 sentences)
• whether it’s already captured in raw/ (from the script output)

4) Ask for approval before capturing/distilling

Ask the user which items (by number) to capture and/or distill next. Do not capture or distill without explicit approval.

5) Capture approved (raw) items

For each approved URL not already captured in raw/:

./scripts/ytraw "<youtube_url>"

If YouTube access is blocked in this environment, ask the user to run ./scripts/ytraw locally and share the resulting raw file path(s).

6) Distill

Use the make-distilled skill on newly captured raw file(s), and ensure each raw file has distilled_refs in front matter (not in the body).

Guardrails

• Never capture/distill without user approval.
• If the pasted Markdown is incomplete or ambiguous, say so and ask for a better export.
• Do not include private/sensitive items in any published note.

Compounding improvement (lightweight)

After you show the triage, ask 1–2 quick questions:

• “Any false hard-skips (something excluded that you’d want to keep)?”
• “Any new hard-skip categories/keywords you want as defaults?”

If the user gives concrete keywords/categories, update .agents/skills/youtube-history-markdown-triage/filters.json in a small, reviewable diff.